Classification of racehorse limb radiographs using deep convolutional neural networks

Area of Science:

• Veterinary Radiology
• Machine Learning in Veterinary Medicine
• Equine Imaging

Background:

• Accurate classification of equine limb radiographs is crucial for veterinary inspections.
• Standardized views are essential for consistent radiological assessment.
• Deep learning offers potential for automating image classification tasks.

Purpose of the Study:

• To evaluate the efficacy of deep convolutional neural networks (CNNs) in classifying anatomical location and projection of equine limb radiographs.
• To assess CNN performance across 48 standard radiographic views.

Main Methods:

• Utilized 9504 equine limb radiographs from 10 veterinary clinics for training, validation, and testing.
• Implemented and evaluated six deep learning architectures from the PyTorch framework.
• Investigated the impact of batch size on the performance of top-performing models.

Main Results:

• Top-1 accuracy across six CNN architectures ranged from 0.737 to 0.841.
• The ResNet-34 architecture achieved the highest top-1 accuracy, ranging from 0.809 to 0.878, with optimal performance at batch size 8.
• Misclassifications were predominantly related to laterality (91.8%), with class activation maps confirming that anatomical features, not markers, influenced model decisions.

Conclusions:

• Deep convolutional neural networks demonstrate capability in classifying equine pre-import radiographs into 48 standard views.
• The models showed moderate discrimination of laterality, independent of side marker presence.
• CNNs show promise for enhancing the efficiency and consistency of equine radiographic interpretation.